This corpus is useful for testing multilingual architectures. One can use a zero-shot setting and test one's architecture on such corpus.

Indeed, one can do the following:

1. Train a model on English-only data using Harry Potter and The Republic books in English.
2. Test a model on non-English data using Harry Potter's book translated to Russian.

1. Harry Potter and the Goblet of Fire, J.K. Rowling - English version (full: 6,192 lines)
2. The Republic, Plato - English version (full: 19,846 lines)
3. Harry Potter and the Cursed Child, J.K. Rowling - Russian version (sample of 435 lines)

One can remove manually the first introduction lines from The Republic, Plato.

We provide a bash file preprocessing data by lowercasing, removing accents and filtering short sentences. One can additionally preprocess data according to their needs.

Install Moses tools for preprocessing. We only used lowercasing and remove accent in practice but one can also tokenize with Moses tools using tokenize.sh in utils folder.

sh intall-tools.sh

We expect one parameter which is the minimum length MIN_LEGNTH. All sentences with a shorter estimated number of tokens than MIN_LENGTH are filtered out. By setting MIN_LENGTH to 0, one can keep the corpus without filtering short sentences.

sh prepare-data-book-corpus.sh MIN_LENGTH

The above will preprocess the corpus by:

1. Lowercasing and removing accents and store that in data/processed e.g. harry-potter.tok.en
2. Filtering out short sentences and store that in data/processed e.g. harry-potter_proc.tok.en

One can add additional steps as desired by modifying prepare-data-book-corpus.sh.